The agent skill

The package ships an agent skill: a folder of instructions in the open Agent Skills format, which Claude Code, Codex, Cursor and some seventy other agents load on demand when a task looks like it needs this library. It exists because the failure mode of an agent meeting 324 algorithms it has never read is not refusal — it is confident invention. A plausible import subpath, a plausible parameter, a plausible field name on the result. The skill replaces each of those guesses with a lookup against the same payload this site is built from.

Install#

Two routes, and they install the same skill.

npx skills add IslamBaraka90/Fintech-Algorithms-Library

The CLI copies the skill into your agent's skills directory — into the current project by default, or -g for every project on the machine. Nothing is installed from npm and the project does not have to depend on the library yet, which makes this the route to take when you want the agent to know what is available before deciding whether to use it.

If the package is already a dependency, the skill is on disk already:

node_modules/fintech-algorithms/skills/fintech-algorithms/

It ships inside the npm tarball, so this copy is version-matched to the docs.json sitting beside it in the package root — the same file the lookup script reads. Point the agent at that directory, or copy it into the agent's skills folder if that is the only place it looks.

What is in it#

• SKILL.md — the router. Four non-negotiable rules, the resolution order below, the archetype table, and the conditions under which to load each reference. It is what the agent reads first and often all it reads.
• references/archetypes.md — the five input shapes, with executed examples, for mapping a user's data onto a signature.
• references/ingestion.md — provider to adapter to validation to computation, and the canonical Trade and Bar shapes.
• references/recipes.md — end-to-end tasks: clean a feed, build bars from ticks, produce a multi-indicator report.
• references/pitfalls.md — failure modes that do not throw. The code runs, a number comes out, and the number is wrong.
• scripts/lookup.mjs — the CLI below.

The references are not loaded up front. SKILL.md names the condition under which each one becomes worth its bytes, which is what keeps a skill covering 324 topics cheap to carry.

The lookup script#

scripts/lookup.mjs has no dependencies and no install step. It reads node_modules/fintech-algorithms/docs.json when the package is installed anywhere above the working directory, and otherwise fetches /reference/payload.json and caches it for a day. So it answers from the version you actually have, and still answers when you have installed nothing at all. FINTECH_DOCS_JSON points it at a specific file.

It accepts a slug (rsi), a catalog id (D07-F03-A01), an import subpath or a docs URL, so anything the agent already has in hand is a valid argument:

node scripts/lookup.mjs search bollinger

2 matches

D07-F04-A03  Bollinger Bands
    import { bollingerBands } from "fintech-algorithms/technical-indicators/volatility-and-channels/bollinger-bands";
    bollingerBands(close, p, multiplier) · series-transform · verified

D07-F04-A06  Bollinger BandWidth
    import { bollingerBandwidth } from "fintech-algorithms/technical-indicators/volatility-and-channels/bollinger-bandwidth";
    bollingerBandwidth(close, p, multiplier) · series-transform · verified

Every part of that came out of the payload: the import path, the parameter names, the archetype, the verification tier. show bollinger-bands goes on to give the parameter contract, the warm-up, the errors it throws and a worked example that was executed rather than typed. When a topic does not exist, the script says so instead of returning the nearest thing — that answer is the correct one, not an obstacle to route around.

What it changes about the answers you get#

• It does not invent a subpath. Every subpath mirrors its docs URL exactly, which is precisely what makes a guess dangerous: it reads as correct. The skill's first rule is to look the path up or say the topic does not exist.
• Warm-up is read, not estimated. show sma reports window - 1 leading nulls and says why they are null rather than partial; show rsi reports p. An agent working from the skill tells you how many leading values are warm-up instead of quietly counting them as signal.
• The verification tier is stated. 158 of 324 topics are verified — the arithmetic is replayed and asserted on every build. The other 166 are contract: the signature and shape are checked, the numbers are not attested by an independent published figure. The skill requires that distinction to be said out loud on any numeric claim.
• It knows the return keys are inconsistent. bollingerBands returns percent_b; a macd row carries fastEma. There is no rule that recovers both, so the skill takes field names from the captured example output and never from the return type.

The boundary#

One rule in the skill is not about correctness. An indicator crossing is an observation about a series, not a recommendation for a specific person's money, and the skill defers to the statement on this site rather than paraphrasing it into something softer: Analysis, not advice.

Which agents can use it#

Any agent that reads the Agent Skills format. The skills CLI installs into around seventy-five of them, and where the files land depends on the agent:

.agents/skills/ is the shared convention, so one project-level install often serves several agents at once. The full list — OpenCode, Copilot, Gemini CLI, Amp, Zed, Goose and the rest — is in the CLI's own documentation.

If your agent reads skills from somewhere else, it is a directory of markdown and one script: copy it there, or read Using this library from an agent, which is the same material written to be fetched rather than installed.